[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Help building Pen.el (GPT for emacs)

From: Shane Mulligan
Subject: Re: Help building Pen.el (GPT for emacs)
Date: Fri, 16 Jul 2021 11:52:41 +1200

Hi Stefan and dick,

* Reponse to Stefan
** Capabilities of "GPT-3+vscode" (Copilot)
Copilot uses a specialised version of GPT-3 called codex which is optimised to generate code.
Copilot is technologically capable of also querying (classifying) code.
The simulateous usage of both AI generation and classification completely changes the way you use an editing environment.

*** Why it's an emacs killer
I must coin a term, imaginary programming, for the sake of shortening my
explanation.  Imaginary programming is imaginary in the mathematical and
creative sense. It's a dimension of programming non-existent in emacs.  It's
stochastic and allows you to predict what will happen without needing to write
correct code.

Clear demonstration here:

- meta-prompts (see below)
- there is no GPT support for emacs.
  - This is the elephant in the room.
  - Asides from what me, a single person in the world is struggling to put
    together. I'm rate-limited. There needs to be core support for integrating
    and designing prompts.
    - I'm working on core features rather than building prompts.
- An analogy: Copilot + GPT-3 + vscode are now firmly in the area of imaginary programming.
    completely missing with emacs. And that's very worrisome.
  - It's the cause of anxiety about what is the future of programming and is it going to make all coders redundant.
    - It's closed source so people are literally unable to see the way ahead.
- It's an attack against open-source.

** Risky capabilities of Copilot in the near future
- meta-prompts that encode the provenance of text. This is an existential risk
  to coding, generally, because right now, today, GitHub (Microsoft) and OpenAI
  (the not-for-profit turned for-profit) are encoding using Copilot the way
  that people are solving problems in order to further automate that process,
  and it's closed-source.
- imaginary generation of user interfaces, such as emacs. There are decades
  worth of text from emacs online and on GitHub. That means GPT-3 and codex are
  quite capable of already imagining at least part of the emacs user interface
  via prompting. The question is, do you want emacs to be in control of your
  system or a copilot that you can't control?

** Capabilities of "Pen.el"
My vision for Pen.el has always been much broader.
- n-many language models
- n-many prompts (classification and generation)
- sharing prompts, open-source
- file format for encoding the provenance of text
  - https://github.com/semiosis/ink.el
- fully transparent
- Use emacs as an interface to remain in control of conversations, whether they
  be infinitely many chatbots or solving communication barriers between people.
  - When I use words such as infinite, I mean it in the truest sense, and it's
    not hype. I've done my best to be a harbinger.
    - Literally, just select any topic and create a chatbot for it.
    - https://semiosis.github.io/posts/gpt-3-for-building-mind-maps-with-an-ai-tutor-for-any-topic/

*** To catch up and surpass and save open-source
- We need a centralised repository of 'prompts', like melpa
  - This doesn't exist yet because the technology is closed-source inside
    GitHub copilot and there are not any standard formats.
  - I have made a start with this
    - https://github.com/semiosis/prompts/
- Pen needs to be integrated (working on this currently)
- GPT-j needs to be integrated (working on this currently)

I have many blog articles now of me trying to demonstrate what the capabilities of Pen are.


* In reponse to dick's message:
"The Copilot brouhaha takes the premature optimization and pointless
speculation endemic to emacs-devel to another level."

I do not believe this is pointless speculation.

On Thu, Jul 15, 2021 at 11:58 PM Stefan Kangas <stefan@marxist.se> wrote:
Shane Mulligan <mullikine@gmail.com> writes:

> Hey guys. It looks like OpenAI is collaborating with GitHub on their GPT stuff, so any assistance in building an editor in emacs would be greatly appreciated. I made a start 4 months ago, link below:
> https://github.com/semiosis/pen.el/
> https://copilot.github.com/
> https://news.ycombinator.com/item?id=27676266
> GPT-3+vscode is an emacs killer, but emacs is a much better platform for building this stuff, so please help! Thanks.

Could you briefly elaborate on the capabilities of "GPT-3+vscode" and
why you think it is an Emacs killer?

What are (briefly) the capabilities of your package so far, and how
does it compare to what the competition has?  How much work remains if
we would want to catch up?


reply via email to

[Prev in Thread] Current Thread [Next in Thread]